Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Even with repeated real exam scores as anchors, local peer talk under incomplete social views can steadily warp a class’s shared sense of who ranks high.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 21:16 UTC pith:2DA3ABV2

load-bearing objection Real-classroom multi-agent sim with a clean DPAE rise under local visibility; useful as mechanism work, but ablations undercut the causal story the abstract sells. the 3 major comments →

arxiv 2603.20750 v2 pith:2DA3ABV2 submitted 2026-03-21 cs.AI

Subjective-Graph LLM Agents for Simulating Uncertainty in Classroom Social Perception

Jinming Yang , Xinyu Jiang , Xinshan Jiao , Xinping Zhang This is my paper
classification cs.AI
keywords multi-agent systemslarge language modelsepistemic uncertaintysocial perceptionsubjective graphscognitive social structuresclassroom networksBayesian belief update
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether students who never share a full view of the classroom network can still drift into a lasting, group-level misreading of academic standing. The authors build a multi-agent simulation in which each student is an LLM agent confined to a personal “subjective graph” derived from questionnaires: that graph alone decides whom they can see, what evidence they can retrieve, and whom they can talk to. Agents swap uncertainty-tagged narrative assessments, score message credibility with the language model, and fuse those signals into explicit Gaussian beliefs via Bayesian updates, while every exam re-anchors each agent on its own score. Across twelve middle-school classrooms and six exams, collective ranking error on the socially observed students roughly doubles—from about 0.066 to 0.124—despite that repeated objective anchoring. Ablations link individualized visibility and LLM trust gating to more stable long-horizon behavior, and the setup keeps more opinion diversity than classical averaging models that collapse toward near-zero disagreement. A sympathetic reader cares because the work offers a concrete, data-constrained way to watch misperception accumulate from local incomplete views rather than from a god’s-eye network.

Core claim

Graph-local evidence and credibility-weighted communication among LLM agents, each confined to an individualized subjective social graph, can generate persistent and accumulating distortions in perceived academic standing even when agents repeatedly receive objective exam scores as self-anchors. On the Social-Observed subset of 419 students in twelve classrooms, group ranking error rises from 0.066 ± 0.008 to 0.124 ± 0.009 over six epochs, while the framework avoids the near-total opinion collapse seen in evaluated DeGroot configurations.

What carries the argument

Individualized subjective graphs: questionnaire-derived personal networks that gate retrieval-augmented evidence, interaction partners, and reachability, combined with LLM trust gating and precision-weighted Bayesian fusion of Gaussian peer-ability beliefs.

Load-bearing premise

The central premise is that language-model assessments and trust scores, run on questionnaire-built personal graphs with parametric noise for social anxiety, are good enough stand-ins for how real students judge peers—without checking simulated peer rankings against students’ own reported rankings of classmates.

What would settle it

Re-run the same six-epoch protocol and check whether group ranking error still rises as reported; or collect independent student-reported peer rankings in the same classrooms and test whether the simulated beliefs track those reports rather than only the exam-score ranking.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Collective misperception of academic rank can grow across successive exams even when every student keeps receiving an objective self-score.
  • Heterogeneous local visibility and explicit trust gating matter more for long-horizon ranking fidelity than unconstrained global retrieval.
  • Classical consensus dynamics that drive opinion diversity near zero can look competitive on correlation while erasing useful disagreement.
  • Classrooms can differ systematically in how fast misperception accumulates, depending on subjective network structure.
  • Language models can be used as structurally constrained cognitive modules rather than omniscient oracles in multi-agent social simulation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the mechanism holds, interventions that expand trusted local visibility may reduce ranking misperception more than simply giving more frequent grades.
  • The same subjective-graph design is a natural stress test for adult workplaces or online cohorts where sparse social data and performance signals co-exist.
  • Comparing simulated peer rankings to students’ actual questionnaire rankings of classmates would most cleanly separate cognitive realism from fitting exam ranks alone.
  • Preserving opinion diversity may be as useful a diagnostic as ranking error when evaluating social simulators.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a multi-agent simulation framework in which LLM agents, each restricted to an individualized subjective social graph built from questionnaire data, exchange uncertainty-annotated peer assessments, apply LLM-derived trust weights, and maintain Gaussian beliefs updated by precision-weighted Bayesian fusion. Using 12 middle-school classrooms (482 students; Social-Observed n=419) and six consecutive exams as exogenous self-anchors, the authors report that collective ranking error (DPAE) rises from 0.066±0.008 to 0.124±0.009 despite repeated score anchoring. Ablations remove constrained RAG, shared vs individualized visibility, and LLM trust; external baselines include feature regressors, simple graph aggregators, and DeGroot dynamics. The stated contribution is mechanism-oriented: showing that graph-local evidence and credibility-weighted communication can sustain structured misperception without global visibility, while preserving more opinion diversity than consensus-collapsing DeGroot runs.

Significance. If the results hold under tighter causal and external checks, the work is a useful methodological contribution at the intersection of multi-agent systems, epistemic uncertainty, and educational social networks. Strengths include real multi-classroom longitudinal scores, explicit probabilistic belief states (Eqs. 2–3), multi-seed reporting with means±std, ablations, DeGroot diversity comparisons, and released code. The framing as a constrained simulation of perception—not a stronger rank predictor—is appropriate. The main scientific value is a concrete architecture for embedding LLMs under local visibility and Bayesian fusion rather than a new empirical law of classroom psychology. That value is reduced until mechanism attribution and external face validity of the simulated perceptions are clarified.

major comments (3)
  1. [§IV.C, Table I; Abstract] The central claim that graph-local evidence and credibility-weighted communication generate the reported DPAE trajectory is only weakly supported by Table I. No-RAG reaches DPAE6=0.121±0.011 and ρ6=0.879, matching or slightly beating Baseline (0.124±0.009, 0.876), while No-Subjective-Graph and No-LLM-Trust produce only modest degradations (0.128 and 0.131). The abstract and §IV.C still credit individualized visibility and LLM trust for “more stable long-horizon behavior” and treat constrained retrieval mainly as an anti-leakage safeguard. That reading is not forced by the numbers: the accumulation of ranking error appears largely driven by repeated noisy narrative messaging plus Gaussian fusion under some visibility, not specifically by subjective-graph-constrained RAG. A load-bearing control is missing—e.g., score-only or frozen-message communication, or a non-LLM observation channel—so
  2. [§III.A–C; §IV.B; Abstract/Title] All evaluation is against exam-induced ground-truth ranks R(t) (Eq. 6 and §IV.A–B). The paper never compares simulated peer rankings or belief means to students’ actual reported rankings or competence judgments of classmates, even though the questionnaire supplies social-relational and psychometric inputs used to build G_i and α_i (§III.A). The title, abstract, and introduction frame the object of study as classroom social perception and subjective academic standing. Under that framing, agreement with exam ranks alone shows that the simulation can drift from objective performance; it does not show that the drift resembles real student misperception. At minimum, the manuscript should either (i) report any available peer-perception items as external validation, or (ii) substantially narrow claims from “social perception” to “simulated ranking beliefs under local interaction,” and discuss t
  3. [§IV.D, Tables II–III; Abstract] The DeGroot comparison in §IV.D / Tables II–III is informative on diversity collapse but incomplete as a ranking baseline. One-step DeGroot achieves DPAE6=0.048 and Acc@3=0.528, and five-step DeGroot still has DPAE6=0.103 with Acc@3=0.444—both better ranking metrics than the proposed method’s Epoch-6 DPAE 0.124 and Acc@3 0.278—while the paper highlights the 30-step configuration (ρ6=0.846, Div6≈5×10−4) as the main foil. The diversity–accuracy trade-off is real and worth reporting, but the claim that the framework “achieves lower final ranking error” than “evaluated DeGroot configurations” (Abstract) depends on which step count is treated as the comparator. Please report the full step sweep as primary baselines, state the selection rule for the DeGroot configuration used in headline comparisons, and qualify “lower final ranking error” accordingly rather than implying a uniform win over De
minor comments (6)
  1. [Title; Abstract; Fig. 1] Title/abstract use “Subjective-Graph LLM Agents…” while the body and Fig. 1 use “Rashomon Set Agents.” Align naming throughout for indexing and clarity.
  2. [§III.C, Eq. (3)] Eq. (3) writes the Bayesian update without time/round superscripts on μ, σ, τ, ŝ; the surrounding text uses (t,r). Make the notation consistent.
  3. [§IV.A Protocol; §III.A] Free parameters R=2, K=3, σ_self, and the mapping from questionnaire scales to α_i are stated but not sensitivity-tested beyond the three ablations. A short appendix sweep (or explicit fixed defaults with justification) would help reproducibility.
  4. [§IV.A Metrics; Table I–II] Fig. 2–8 captions are clear, but Acc@3 is reported without a chance baseline or class-size normalization; with varying classroom sizes, raw Top-3 accuracy is hard to interpret across classes.
  5. [§II Related Work] Related Work cites DeGroot, cognitive social structures (Krackhardt), and LLM multi-agent surveys appropriately; a brief pointer to empirical classroom peer-perception / sociometric ranking studies would better situate the missing external validation.
  6. [Abstract; throughout] Minor prose issues: “W e” / “T o” spacing artifacts in the abstract; “difficult”/“insufficiently” ligature rendering; “Full-Temporal” hyphenation inconsistency.

Circularity Check

0 steps flagged

No circular derivation: rising DPAE is a simulation outcome under external exam anchors and baselines, not an identity or fitted-parameter rename.

full rationale

The paper is a mechanism-oriented multi-agent simulation, not a first-principles derivation of a closed-form prediction. Exam scores Y^(t) enter only as exogenous self-anchors (Eq. 4) and as the evaluation target R^(t) for DPAE = 1 − Spearman(ˆR, R); peer beliefs are Gaussian states updated by LLM messages and trust-weighted Bayesian fusion (Eqs. 2–3) under questionnaire-built subjective graphs, not by fitting parameters to the DPAE trajectory. Nothing in the update equations forces DPAE to rise—self-precision is high, and DeGroot/regressor/GNN baselines provide independent comparisons that can and do behave differently (e.g., DeGroot consensus collapse). Ablations and external baselines further show the reported numbers are empirical run outputs, not tautologies of the inputs. There is no self-citation uniqueness theorem, no ansatz smuggled from the authors’ prior work as a forced law, and no renaming of a known closed result as a new derivation. Designing a model that can accumulate local noise and then observing accumulation under real classroom constraints is standard simulation practice, not circularity under the stated criteria.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central claim rests on modeling choices that turn questionnaire fragments into private graphs and LLM outputs into trust-weighted Gaussian updates, then measures misperception only against exam ranks. Free parameters (rounds, partners, self-variance, anxiety noise) and invented constructs (subjective graphs, trust-gated precision) do most of the work; standard Bayesian Gaussian fusion and DeGroot comparisons are background. Without independent student ranking labels, the ledger is heavy on domain/ad-hoc modeling assumptions relative to what is measured.

free parameters (6)
  • interaction_rounds_R
    Fixed R=2 social rounds per exam epoch; controls how much local noise can accumulate between anchors.
  • partners_per_round_K
    Fixed K=3 sampled interaction partners from subjective neighborhood; shapes message volume and diffusion.
  • sigma_self
    Self-belief variance after exam anchor set to a small constant σ_self ≪ 1; strength of objective self-calibration is hand-chosen.
  • alpha_i_social_anxiety_noise
    Per-agent noise mapped from questionnaire scales controlling missing/spurious edges and retrieval perturbation; mapping function and scale are paper-specific and not independently validated.
  • LLM_trust_weight_omega
    ω_i→j ∈ [0,1] produced by an LLM from constrained evidence and attributes; effectively a free, model-dependent credibility map each message.
  • observation_precision_tau_from_omega_and_u
    Message precision τ constructed from trust and self-reported uncertainty; functional form modulates Bayesian fusion strength and is design-chosen.
axioms (6)
  • domain assumption Beliefs about peer ability are well-approximated by independent univariate Gaussians B_jk = N(μ_jk, σ_jk).
    Stated in §III.C for tractability; real social judgments may be multimodal, discrete ranks, or correlated across peers.
  • standard math Message influence is precision-weighted Bayesian fusion of Gaussian prior and scalar observation.
    Eq. (3) is standard conjugate Gaussian update; validity depends on treating LLM messages as calibrated Gaussian observations.
  • domain assumption Questionnaire friendships and psychometric scales suffice to build individualized subjective graphs G_i and reachability weights π_i.
    §III.A construction; assumes sparse self-report captures the visibility that drives classroom perception.
  • ad hoc to paper LLM narrative assessments under RAG locality are valid stand-ins for student communication and credibility judgment.
    Core of §§III.B–C; no human-message or human-trust calibration is reported.
  • domain assumption Group-perceived ranking from averaged belief means is the right collective observable for ‘social perception’ of academic standing.
    Eqs. (5)–(6); equates simulation aggregate to social perception without student-reported ranks.
  • domain assumption Exam score map h(Y) places ability on a shared scale usable for cross-student belief fusion.
    §III.C self-calibration; assumes scores are commensurable ability signals across students and time.
invented entities (3)
  • Individualized subjective social graph G_i = (V, E_i, π_i) no independent evidence
    purpose: Restrict each agent’s visibility, retrieval, and interaction partners away from a global network.
    Builds on cognitive social structures but is operationalized here as the load-bearing locality mechanism with anxiety noise α_i.
  • Rashomon Set Agents / subjective-graph LLM agent population no independent evidence
    purpose: Name the multi-agent system that couples constrained RAG, trust gating, and Gaussian belief states.
    Framework-level construct; evidence is internal simulation behavior, not an external measured entity.
  • LLM trust-gating weight ω and induced observation precision τ no independent evidence
    purpose: Filter noisy peer messages before Bayesian fusion.
    New operational module in this pipeline; no external ground-truth credibility labels.

pith-pipeline@v1.1.0-grok45 · 16586 in / 3899 out tokens · 42893 ms · 2026-07-13T21:16:31.254286+00:00 · methodology

0 comments
read the original abstract

Social actors do not observe a common social world: each individual forms judgments from a partial and potentially distorted view of the surrounding network. We study whether graph-local evidence and credibility-weighted communication can generate persistent distortions in perceived academic standing, even when agents repeatedly receive objective performance signals. We introduce a data-constrained multi-agent framework in which LLM agents operate through individualized subjective graphs that determine peer visibility, evidence access, and interaction opportunities. Agents exchange uncertainty-annotated assessments, evaluate message credibility, and maintain explicit Gaussian belief states updated through Bayesian fusion. We evaluate the framework on 12 middle-school classrooms comprising 482 students, using questionnaire-derived social information and six consecutive examinations. On the Social-Observed subset (n=419), collective ranking error increases from 0.066 \pm 0.008 to 0.124 \pm 0.009 across six epochs despite repeated exam-based anchoring. Ablations associate individualized visibility and LLM-based trust gating with more stable long-horizon behavior, while constrained retrieval primarily safeguards against global-information leakage. Compared with evaluated DeGroot configurations, the proposed framework achieves lower final ranking error; those DeGroot configurations exhibit near-zero terminal opinion diversity. These findings establish subjective-graph LLM agents as a mechanism-oriented framework for data-constrained simulated social perception. Code is available at https://anonymous.4open.science/r/Rashomonomon-0126.

Figures

Figures reproduced from arXiv: 2603.20750 by Jinming Yang, Xinping Zhang, Xinshan Jiao, Xinyu Jiang.

Figure 1
Figure 1. Figure 1: Overview of Rashomon Set Agents: (A) subjective graphs and retrievable knowledge; (B) graph-constrained RAG messaging with LLM trust gating and Bayesian updates; (C) multi-epoch outputs and diagnostics. III. Method Consider a classroom system composed of N students, with node set V = {1, . . . , N}. The system evolves on a discrete timeline t ∈ {1, . . . , 6}, where each time step corresponds to one real e… view at source ↗
Figure 2
Figure 2. Figure 2: Baseline temporal trend: DPAE increases and Spearman [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Paired ablation effects vs. Baseline under shared seeds [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Epoch 6 comparison: Baseline vs. ablations on DPAE, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: DeGroot sweep at Epoch 6: more steps increase consensus [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Class-wise heterogeneity at Epoch 6: final DPAE varies [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 2 canonical work pages

  1. [1]

    Educational data mining to predict bachelors students’ success,

    D. Jacob and R. Henriques, “Educational data mining to predict bachelors students’ success,” Emerging Science Journal , vol. 7, pp. 159–171, 2023

    2023

  2. [2]

    Systematic review of research on artificial intelligence applica- tions in higher education—where are the educators?

    O. Zawacki-Richter, V. I. Marín, M. Bond, and F. Gouverneur, “Systematic review of research on artificial intelligence applica- tions in higher education—where are the educators?” Interna- tional Journal of Educational Technology in Higher Education , vol. 16, no. 1, p. 39, 2019

    2019

  3. [3]

    It’s who you know: A review of peer networks and academic achievement in schools,

    A. Black, M. F. Warstadt, and C. Mamas, “It’s who you know: A review of peer networks and academic achievement in schools,” Frontiers in Psychology, vol. 15, p. 1444570, 2025, doi: 10.3389/fpsyg.2024.1444570

    work page doi:10.3389/fpsyg.2024.1444570 2025

  4. [4]

    Students’ academic performance prediction using educational data mining and machine learning: A system- atic review,

    M. S. N. Al-Din, “Students’ academic performance prediction using educational data mining and machine learning: A system- atic review,” International Journal of Research and Innovation in Social Science , vol. 8, no. 8, pp. 1264–1291, 2024

    2024

  5. [5]

    Multi-agent reinforcement learning: A comprehensive survey,

    D. Huh and P. Mohapatra, “Multi-agent reinforcement learning: A comprehensive survey,” arXiv preprint arXiv:2312.10256 , 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Large language models for agent-based mod- elling: Current and possible uses across the modelling cycle,

    L. Vanhée, M. Borit, P.-O. Siebers, R. Cremades, C. Frantz, Ö. Gürcan, F. Kalvas, D. R. Kera, V. Nallur, K. Narasimhan, and M. Neumann, “Large language models for agent-based mod- elling: Current and possible uses across the modelling cycle,” arXiv preprint arXiv:2507.05723 , 2025

    Pith/arXiv arXiv 2025

  7. [7]

    A survey on large language model based autonomous agents,

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y. Lin, W. X. Zhao, Z. Wei, and J.-R. Wen, “A survey on large language model based autonomous agents,” Frontiers of Computer Science , vol. 18, no. 6, p. 186345, 2024

    2024

  8. [8]

    Large language model based multi- agents: A survey of progress and challenges,

    T. Guo, X. Chen, Y. Wang, R. Chang, S. Pei, N. V. Chawla, O. Wiest, and X. Zhang, “Large language model based multi- agents: A survey of progress and challenges,” in Proc. 33rd Int. Joint Conf. Artificial Intelligence (IJCAI), 2024, pp. 8048–8057

    2024

  9. [9]

    Beyond static responses: Multi-agent LLM systems as a new paradigm for social science research,

    J. Haase and S. Pokutta, “Beyond static responses: Multi-agent LLM systems as a new paradigm for social science research,” arXiv preprint arXiv:2506.01839 , 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Graph neural networks for social recommendation,

    W. Fan, Y. Ma, Q. Li, Y. He, E. Zhao, J. Tang, and D. Yin, “Graph neural networks for social recommendation,” in Proc. World Wide Web Conf. (WWW) , 2019, pp. 417–426

    2019

  11. [11]

    INFLECT-DGNN: Influencer pre- diction with dynamic graph neural networks,

    E. Tiukhova, E. Penaloza, M. Óskarsdóttir, B. Baesens, M. Snoeck, and C. Bravo, “INFLECT-DGNN: Influencer pre- diction with dynamic graph neural networks,” IEEE Access , vol. 12, pp. 115 026–115 041, 2024

    2024

  12. [12]

    AI-based conversational agents: A scoping review from technologies to future directions,

    S. Kusal, S. Patil, J. Choudrie, K. Kotecha, S. Mishra, and A. Abraham, “AI-based conversational agents: A scoping review from technologies to future directions,” IEEE Access, vol. 10, pp. 92 337–92 356, 2022

    2022

  13. [13]

    LLMs and generative agent-based models for complex systems research,

    Y. Lu, A. Aleta, C. Du, L. Shi, and Y. Moreno, “LLMs and generative agent-based models for complex systems research,” Physics of Life Reviews , vol. 51, pp. 283–293, 2024

    2024

  14. [14]

    Population-aligned persona generation for LLM-based social simulation,

    Z. Hu, J. Lian, Z. Xiao, M. Xiong, Y. Lei, T. Wang, K. Ding, Z. Xiao, N. J. Yuan, and X. Xie, “Population-aligned persona generation for LLM-based social simulation,” arXiv preprint arXiv:2509.10127, 2025

    arXiv 2025

  15. [15]

    Reaching a consensus,

    M. H. DeGroot, “Reaching a consensus,” Journal of the Ameri- can Statistical Association, vol. 69, no. 345, pp. 118–121, 1974

    1974

  16. [16]

    Social networks in structural equation models,

    N. E. Friedkin, “Social networks in structural equation models,” Social Psychology Quarterly , vol. 53, no. 4, pp. 316–328, 1990, doi: 10.2307/2786737

    work page doi:10.2307/2786737 1990

  17. [17]

    Continuous opinion dynamics under bounded confi- dence: A survey,

    J. Lorenz, “Continuous opinion dynamics under bounded confi- dence: A survey,” International Journal of Modern Physics C , vol. 18, no. 12, pp. 1819–1838, 2007

    2007

  18. [18]

    Naive learning in social networks and the wisdom of crowds,

    B. Golub and M. O. Jackson, “Naive learning in social networks and the wisdom of crowds,” American Economic Journal: Mi- croeconomics, vol. 2, no. 1, pp. 112–149, 2010

    2010

  19. [19]

    Agent-based modelling meets generative AI in social network simulations,

    A. Ferraro, A. Galli, V. La Gatta, M. Postiglione, G. M. Orlando, D. Russo, G. Riccio, A. Romano, and V. Moscato, “Agent-based modelling meets generative AI in social network simulations,” arXiv preprint arXiv:2411.16031 , 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Spread of (mis)information in social networks,

    D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi, “Spread of (mis)information in social networks,” Games and Economic Behavior, vol. 70, no. 2, pp. 194–227, 2011

    2011

  21. [21]

    Model- ing interpersonal perception in dyadic interactions: Towards robot-assisted social mediation in the real world,

    H. Javed, W. Wang, A. B. Usman, and N. Jamali, “Model- ing interpersonal perception in dyadic interactions: Towards robot-assisted social mediation in the real world,” Frontiers in Robotics and AI , vol. 11, p. 1410957, 2024

    2024

  22. [22]

    Neural relational inference for interacting systems,

    T. N. Kipf, E. Fetaya, K.-C. Wang, M. Welling, and R. Zemel, “Neural relational inference for interacting systems,” in Proc. 35th Int. Conf. Machine Learning (ICML) , 2018

    2018

  23. [23]

    Relational inductive biases, deep learning, and graph networks,

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. San- toro, R. Faulkner, C. Gulcehre, F. Song, A. Ballard, J. Gilmer, G. E. Dahl, A. Vaswani, K. Allen, C. Nash, V. Langston, C. Dyer, N. Heess, D. Wierstra, P. Kohli, M. Botvinick, O. Vinyals, L. Li, and R. Pascanu, “Relational inductive ...

    Pith/arXiv arXiv 2018

  24. [24]

    GNN-IR: Examining graph neural networks for influencer recommendations in social media marketing,

    J. Park, H. Ahn, D. Kim, and E. Park, “GNN-IR: Examining graph neural networks for influencer recommendations in social media marketing,” Journal of Retailing and Consumer Services , vol. 78, p. 103759, 2024

    2024

  25. [25]

    Generative agents: Interactive simulacra of human behavior,

    J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” in Proc. 36th Annu. ACM Symp. User Interface Software Technol. (UIST) , 2023

    2023

  26. [26]

    Bayesian teaching enables prob- abilistic reasoning in large language models,

    L. Qiu, F. Sha, and K. Allen, “Bayesian teaching enables prob- abilistic reasoning in large language models,” Nature Commu- nications, 2026

    2026

  27. [27]

    BayesAgent: Bayesian agentic rea- soning under uncertainty via verbalized probabilistic graphical modeling,

    H. Huang, X. Shen, S. Wang, L. Meng, D. Liu, D. A. Duchene, H. Wang, and S. Bhatt, “BayesAgent: Bayesian agentic rea- soning under uncertainty via verbalized probabilistic graphical modeling,” arXiv preprint arXiv:2406.05516 , 2024

    arXiv 2024

  28. [28]

    Are LLM belief updates consistent with Bayes’ theorem?

    S. Imran, I. Kendiukhov, M. Broerman, A. Thomas, R. Cam- panella, R. Lamb, and P. M. Atkinson, “Are LLM belief updates consistent with Bayes’ theorem?” in Proc. ICML Workshop on World Models, 2025

    2025

  29. [29]

    Uncer- tainty propagation on LLM agent,

    Q. Zhao, D. Li, Y. Liu, W. Cheng, Y. Sun, M. Oishi, T. Osaki, K. Matsuda, H. Yao, C. Zhao, H. Chen, and X. Zhao, “Uncer- tainty propagation on LLM agent,” in Proc. 63rd Annu. Meeting Assoc. Comput. Linguist. (ACL) , 2025, pp. 6064–6073

    2025

  30. [30]

    The emerging field of emotion regulation: An integrative review,

    J. J. Gross, “The emerging field of emotion regulation: An integrative review,” Review of General Psychology , vol. 2, no. 3, pp. 271–299, 1998

    1998

  31. [31]

    R. W. Picard, Affective Computing . Cambridge, MA, USA: MIT Press, 1997

    1997

  32. [32]

    Large language models show amplified cognitive biases in moral decision-making,

    V. Cheung, M. Maier, and F. Lieder, “Large language models show amplified cognitive biases in moral decision-making,” Pro- ceedings of the National Academy of Sciences , vol. 122, no. 22, p. e2412015122, 2025

    2025

  33. [33]

    Mitigating cognitive biases in clinical decision-making through multi-agent conversa- tions using large language models: Simulation study,

    Y. Ke, R. Yang, S. A. Lie, T. X. Y. Lim, Y. Ning, I. Li, H. R. Abdullah, D. S. W. Ting, and N. Liu, “Mitigating cognitive biases in clinical decision-making through multi-agent conversa- tions using large language models: Simulation study,” Journal of Medical Internet Research , vol. 26, p. e59439, 2024

    2024

  34. [34]

    LLM agent-based simulation of student activities and mental health using smartphone sensing data,

    W. Sommuang, K. Kerdthaisong, P. Buakhaw, A. B. Wong, and N. Yongsatianchot, “LLM agent-based simulation of student activities and mental health using smartphone sensing data,” arXiv preprint arXiv:2508.02679 , 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Explainable multi-agent reinforcement learning for temporal queries,

    K. Boggess, S. Kraus, and L. Feng, “Explainable multi-agent reinforcement learning for temporal queries,” in Proc. 32nd Int. Joint Conf. Artificial Intelligence (IJCAI) , 2023, pp. 55–63

    2023

  36. [36]

    An empirical evaluation of the rashomon effect in ex- plainable machine learning,

    S. Müller, “An empirical evaluation of the rashomon effect in ex- plainable machine learning,” arXiv preprint arXiv:2306.15786 , 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Cognitive social structures,

    D. Krackhardt, “Cognitive social structures,” Social Networks , vol. 9, no. 2, pp. 109–134, 1987

    1987

  38. [38]

    Auto-RAG: Autonomous retrieval-augmented generation for large language models,

    T. Yu, S. Zhang, and Y. Feng, “Auto-RAG: Autonomous retrieval-augmented generation for large language models,” arXiv preprint arXiv:2411.19443 , 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Retrieval-augmented gen- eration for large language models: A survey,

    Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, M. Wang, and H. Wang, “Retrieval-augmented gen- eration for large language models: A survey,” arXiv preprint arXiv:2312.10997, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Attention knows whom to trust: Attention-based trust management for LLM multi-agent systems,

    P. He, Z. Dai, X. Tang, Y. Xing, H. Liu, J. Zeng, Q. Peng, S. Agrawal, S. Varshney, S. Wang, J. Tang, and Q. He, “Attention knows whom to trust: Attention-based trust management for LLM multi-agent systems,” arXiv preprint arXiv:2506.02546, 2025

    Pith/arXiv arXiv 2025